Artificial intelligence is shaking up the pharmaceutical industry, delivering efficiencies in manufacturing, quality control, and real-time monitoring. But with great potential comes great responsibility—especially in regulated environments. For us IT & Quality directors, ensuring patient safety, product quality, and data integrity is not negotiable. The challenge? AI is moving faster than compliance standards, leaving many unsure of what they need to do.
AI models don’t follow predefined logic like traditional software. They’re non-deterministic, operating on probabilities, which presents a regulatory puzzle. Traditional validation frameworks like GAMP 5, and requirements such as 21 CFR Part 11, and Annex 11 can still apply. Impact assessments, business requirements, and risk evaluations are all still relevant however traditional positive, negative and boundary testing often used as forms of risk mitigation may no longer be appropriate.
AI throws in new concepts to consider, like bias assessments, statistical error rates, and monitoring for false positives and negatives. When performing quality assurance on an AI model, it raises new questions for us:
-
How well does the model predict quality, and is it actually better than a non-ML approach?
-
What is the statistical error rate or what statistical documentation can you provide from evaluation?
-
What are the model limitations?
-
What data has it been trained/tested/validated on?
-
Is the data of high quality?
-
Who labelled the data?
-
Has the data been checked for biases, including societal inequities (think clinical trial representation)?
-
How are outliers handled?
-
What are the consequences of false positives / false negatives?
-
Does the model use static or dynamic training?
-
How do you feed deviations back into the model?
-
What is the business use and business risk?
The ability to assess AI quality depends on how well the model’s probabilistic nature is communicated and its impact on product quality and patient safety understood. Without a structured evaluation, companies risk deploying AI systems that lack transparency and explainability—two non-negotiables in regulated environments.
Datasheets and labels are already a standard way to document warnings and limitations in pharma. Model cards[1] apply the same logic to AI, providing a streamlined way to answer critical compliance questions. Here’s a model card specifically designed to support fast QA and compliance feedback:
Model cards give a quick yet comprehensive view of a model’s intended use, limitations, and risks. They ensure regulatory teams have the details they need to evaluate AI’s impact. More than just documentation, model cards establish transparency in decision-making and act as a single source of truth throughout the model’s lifecycle. They empower IT & Quality directors to make informed risk assessments without getting lost in technical jargon.
First and foremost, the model card bridges the chasm between technical engineering and quality assurance for an effective determination of quality & decision-making. Where AI outputs have a direct and critical contribution to a regulatory submission, including AI model cards are an effective way to anticipate and address any regulatory concerns and to reduce any delays. Simplifying the purpose, function, and risk of a model into a structured format helps satisfy what regulators are looking for: clarity, explainability, and traceability.
Bridging the gap between technical and quality with model cards ultimately turns machine learning complexity into actionable insights. They enable risk-based decision-making and instill best practices in AI governance, ensuring models stay auditable, explainable, and GxP-compliant.
Model cards make compliance easier, offering a structured yet flexible way to evaluate an AI model’s reliability, safety, and compliance at a glance. To simplify AI compliance, ERA Sciences has developed a ready-to-use model card template aligned with industry best practices. Download your template today and ensure AI quality assurance is as structured and efficient as the rest of your GxP systems.
References:
-
Model Cards for Model Reporting, Mitchell, M et al (2019)
Comments